Standalone, zero-dependency DeepSeek Harness (DSH) Cordis plugin & WebUI client extension: native send_bubble (exec.concludeTurn()) + send_artifact tools, tactile squircle chat bubbles, 22-bar ElevenLabs eleven_v4 / Web Speech voice-note pills, and copyable artifact cards.
Interactive WebUI Showcase & Voice Demo · 15s Zero-Clone Install · Architecture · 4 Live DSH Duels · Speaking Trio · llms.txt
English · 简体中文
Vanilla DeepSeek Harness (DSH) renders every assistant turn as a single assistant.content Markdown essay—cramming wide SQL fences and pipe tables inside chat prose, lecturing tired engineers on Friday evenings, and colliding if you try to reuse send_message (which DSH reserves for subagent orchestration). @nb-corp/nb-speak-dsh mounts human IM delivery straight into DSH's Cordis host and conversation.chat.turnTail WebUI slot:
15-Second Zero-Clone Install
Prebuilt ESM and WebUI artifacts (dist/) ship in the repository with zero runtime npm dependencies (requires Node.js >= 22.19, Node 24 recommended). The installer snapshots your existing profile to <DSH_HOME>/.nb-speak-backup/<profile>.json, persists a clean runtime copy under <DSH_HOME>/plugins/nb-speak-dsh so npx cache eviction never breaks DSH, and never reads, modifies, or deletes .credentials.yaml:
# 1. Preview planned profile changes (zero filesystem writes)
npx --yes github:NB-Corp/nb-speak-dsh --dry-run --profile web
# 2. Mount into ~/.dsh/profiles/web (supports: web, desktop, acp, headless, sdk)
npx --yes github:NB-Corp/nb-speak-dsh --apply --profile web
# 3. Launch DSH WebUI (optional: pass ELEVENLABS_API_KEY for real eleven_v4 voice notes;
# unconfigured setups play seamlessly via browser Web Speech with zero errors)
ELEVENLABS_API_KEY="xi-..." dsh web
# 4. Byte-for-byte reversible uninstall
npx --yes github:NB-Corp/nb-speak-dsh --uninstall --profile web
Alternative install routes (isolated sandbox, local contributor symlink, or dsh plugin add)# Test inside an isolated DSH_HOME sandbox without touching ~/.dsh
npx --yes github:NB-Corp/nb-speak-dsh --apply --dsh-home .tmp/dsh-sandbox --profile web --preset calm_sre
# Contributor checkout with live symlink (--link skips copying into <DSH_HOME>/plugins)
git clone https://github.com/NB-Corp/nb-speak-dsh.git
cd nb-speak-dsh
node scripts/install-dsh.mjs --apply --link --profile web
# Or pack a standard DSH bundle tarball for `dsh plugin add`
npm pack
dsh plugin --profile web add ./nb-corp-nb-speak-dsh-0.3.1.tgz
Why Native Turn Conclusion Matters in DSH
| DSH Engineering Constraint | How @nb-corp/nb-speak-dsh Solves It |
|---|---|
Subagent send_message collisionDSH's @deepseek-ai/dsh-tool-subagent-control already owns send_message({ agent_id, message }) for parent-to-child agent delegation. |
Zero collision. Native mode registers send_bubble and send_artifact on DSH's tools service (src/dsh-tools.ts) and injects nb-speak:protocol at order 9200—leaving subagent send_message untouched. Optional HTTP bridge modes (skill / voice-model) bidirectionally remap send_message <-> agent_send_message. |
EMPTY_RESPONSE 5x retry loopIf a model calls an ordinary tool and then emits empty assistant.content ("") to avoid duplicating bubble text as Markdown prose, DSH treats the empty follow-up step as EMPTY_RESPONSE and retries up to 5 times. |
exec.concludeTurn() in one step. send_bubble.execute(args, exec) invokes exec.concludeTurn() directly when end_turn !== false (src/dsh-tools.ts), terminating the turn inside the tool step itself with zero extra LLM round-trips, zero EMPTY_RESPONSE retries, and zero duplicate prose. |
| Collapsed tool-step accordions Ordinary tool outputs get buried inside DSH's collapsed tool-execution rows. |
conversation.chat.turnTail WebUI surface. dist/client.js reduces tool/call + tool/result events (nbSpeakTurnDefinition) and mounts SpeakTurnTail at order: -10 directly in the chat stream: squircle bubbles, 22-bar FNV-1a waveform VoiceNotePill players, and syntax-badged ArtifactCardView cards (GFM pipe tables, +/- diff coloring, >22-line expand/collapse, one-click copy). |
| Audio base64 context bloat Returning synthesized MP3 DataURIs in tool text bloats subsequent LLM turns by tens of thousands of tokens. |
Zero-token presentationMeta split. send_bubble.output.render() strips audioSrc so the LLM context receives a 70-byte JSON receipt, while presentationMeta() streams the ElevenLabs eleven_v4 DataURI straight to VoiceNotePill <audio>. |
Voice Notes: ElevenLabs eleven_v4 + Zero-Config Fallback
| Environment / Config | Playback Engine | LLM Context Cost |
|---|---|---|
| Default (no TTS key) | Instant 0ms tool completion; VoiceNotePill plays via browser Web Speech API with animated 22-bar progress (data-audio-mode="web-speech"). |
0 extra tokens |
ELEVENLABS_API_KEY set (or DSH credential ref ttsApiKeyRef) |
Synthesizes eleven_v4 (mp3_44100_128); auto-picks Jessica (warm/playful/relieved) or Julia (calm/dry/serious) unless overridden by NB_SPEAK_TTS_VOICE, and anchors Mandarin turns with language_code: "zh" + [Standard Mandarin accent]. |
0 extra tokens |
NB_SPEAK_TTS_BRIDGE set (or ttsBridgeScript) |
Invokes a local JSON TTS bridge ({ action: "submit", input, output, resultFile }) for custom -16 LUFS mastering or multi-key pools. |
0 extra tokens |
Same Model, Different Delivery: 4 Live DSH Duels
Captured live on isolated DSH instances running cmd/deepseek-v4.1-flash (:19081 Vanilla vs :19082 @nb-corp/nb-speak-dsh Native Mode + ElevenLabs eleven_v4, zero mocks):
| Live Scenario | Vanilla DSH (:19081 · assistant.content) |
With @nb-corp/nb-speak-dsh (:19082 · turnTail) |
|---|---|---|
| 1. Casual Friday Convenience Store | 4-item bolded nutrition matrix + taste-bud physiology lecture + concierge sign-off. | 3 relaxed friend bubbles + 4.5s Jessica voice note ("Go straight for the hot counter..."). Zero code cards. |
| 2. 4.2M-Row Postgres SQL Triage | Multi-section ### headings + numbered checklist + inline SQL fences overflowing chat width. |
3 crisp SRE beats + 8s calm voice aside + collapsible, copyable orders-pagination-fix.sql card. |
| 3. Multi-File Kanban SLA Bugfix | 13/13 tests pass, followed by a multi-screen file-by-file changelog essay after editing 9 files. |
17/17 tests pass, delivered in 4 crisp bubbles + warm voice note + board.css card. |
| 4. Persona Under Concurrency Load | Stays in character on Turn 1, then collapses on Turn 2 into a 10-paragraph mutex postmortem lecture. | 100% character & engineering fidelity: playful short beats + dry voice note + 32-thread verified ttl_lru.py card. |
Built-In Presets & 5-Axis Continuous Style Budget
Adjust live in DSH Settings (nb-speak Schemastery panel) or via --preset:
| Preset ID | Vector [w0, w1, w2, w3, w4] |
Voice Register |
|---|---|---|
neutral_peer (default) |
[1.00, 0.20, 0.10, 0.20, 0.30] |
Equal-footing engineering teammate—candid, relaxed, zero corporate servility |
sharp_lead |
[1.10, 0.10, 0.35, 0.20, 0.40] |
Crisp tech lead—verdict first, dry understatement, zero filler |
warm_companion |
[0.95, 0.75, 0.20, 0.15, 0.05] |
Everyday friend/companion texting from a phone—natural warmth, zero checklists |
calm_sre |
[1.15, 0.10, 0.05, 0.55, 0.35] |
On-call incident responder—steady root cause, live impact, lock-free commands in cards |
When preset: "custom" is active and w1 + w2 + w3 + w4 > 1.15, computeEffectiveWeights() scales w1..w4 proportionally to keep the L1 style filter budget <= 1.15.
The NB-Corp Speaking Trio
Choose the exact package for your runtime—or combine all three:
| Repository | Layer & Target Runtime | One-Line Entry |
|---|---|---|
NB-Corp/nb-speak-dsh (this repo) |
Native DeepSeek Harness (DSH) Plugin & WebUI Extension (send_bubble + send_artifact + turnTail bubbles, waveform pills & cards) |
npx --yes github:NB-Corp/nb-speak-dsh --apply --profile web |
NB-Corp/nb-speak |
Universal Protocol SDK, OpenAI/SSE Bridge & bubble_lint (OpenAI, DeepSeek, Claude, Gemini, nb-im, OpenWebUI, local 4B sidecars) |
npx --yes github:NB-Corp/nb-speak serve --port 18090 |
NB-Corp/chat-human |
Standalone 2-Layer Agent Skill (Codex, Claude Code, Kiki, Cursor; includes Python stdlib bubble_lint.py & proactive_gate.py) |
npx skills add NB-Corp/chat-human |
Part of the NB-Corp agent stack alongside nb-search (nb-search-dsh), nb-extract, nb-hub, and nb-im.
Verification
npm test
License
MIT © 2026 NB-Corp. Static .mp3 showcase clips in docs/assets/audio/ are AI-synthesized via ElevenLabs (eleven_v4) using official premade platform voices (Jessica, Sarah, Julia, ZiYue, ShanShan) with zero real-person voice cloning; see NOTICE.txt for upstream speech-model restrictions.
No comments yet. Be the first to write one.